AI is a great way to make the challenging or tedious aspects of pathology simpler, but pathologists will always be the experts
By Dr. Juan A. Retamero
From the beginning, Paige has been on a mission to support pathologists in the important work we do to care for cancer patients. Each decision that went into the creation of Paige’s AI was driven by this mission, from training our models on one of the largest and most diverse datasets in pathology AI, to using multiple instance to ensure the models are as robust and impervious to pre-analytical variations as possible. For pathologists to best implement our AI into clinical practice, it is important to understand these decisions, so we can trust the AI enough to work in harmony with it. Because when pathologists and AI join forces, we’re stronger together than either alone, and patients reap the benefits.
To understand how our AI works, let’s start at the beginning: Each of Paige’s applications are tools for assisting pathologists in the detection of areas of tissue suspicious for cancer. Therefore, like a diagnostic test, a compromise between sensitivity and specificity has to be reached. A cutoff point that determines the result has to be chosen, which in our case, is the point at which a tissue sample is deemed suspicious for cancer or not. Determining this cutoff point is a challenge for all diagnostic tests, as there can often be a subset of patients whose diagnosis falls within the gray area, and depending on where the threshold is, these patients might be wrongly classified as positives or negatives.
Paige’s goal when designing our AI is to prioritize sensitivity to reduce the likelihood of false negatives as much as possible. That is, by increasing sensitivity, we aim to reduce the likelihood that a patient who may be positive for cancer would be overlooked. As such, each of our products is optimized to have the highest possible sensitivity, including the entirety of the Paige Prostate Suite, which contains tools for detecting prostate cancer, grading and quantifying tumors, identifying perineural invasion (PNI), and uncovering prostate biomarkers.
Paige Prostate Grade & Quantify, our AI application built for categorizing areas suspicious for cancer by Gleason pattern, is a product where optimizing for greater sensitivity* was especially important. The difference between a pattern 3 or 4 prostate cancer tumor, for example, means the difference between non-clinically significant prostate cancer that will be relatively inconsequential for years or even decades, or clinically significant prostate cancer that could go on to invade neighboring organs and potentially cause much more serious outcomes. Gleason grading can also be challenging and highly subjective, leaving room for error when manually determining which cases are clinically significant. Armed with a tool that is able to distinguish Gleason patterns and PNI even in challenging samples, pathologists can increase their confidence that the right patients will receive the right treatment.
Of course, choosing to favor sensitivity does mean that specificity*, while still very high, is reduced slightly. This may cause false positives, or clinically insignificant cases being flagged as significant. For this reason, AI can never replace a human pathologist. With a glance, the expert pathologist can rapidly and easily recognize these cases as false positives and make the correct final diagnosis.
Pathologists then should consider AI as an assistant, just like a resident or fellow can be. Its strength lies in its ability to take the challenging or tedious aspects of pathology and make them simpler and more efficient. For example, pathologists can use AI to prioritize cases in their worklist, reviewing those cases that are most likely suspicious for cancer earlier in the day when they are the most focused. They can also employ it as a second read tool, providing an added layer of confidence in their own expert diagnosis, and reducing the need for external second opinions. This can further help reduce inter- and intra-observer subjectivity, reduce the margin of error , and streamline diagnostic workflows to reduce turnaround times for patients.
In fact, this was the case at The Institute of Molecular Pathology and Immunology of the University of Porto (IPATIMUP), in Porto, Portugal, where a recent independent study of Paige Prostate Detect and Paige Prostate Grade & Quantify was conducted in a real-world clinical setting. The study followed 4 generalist pathologists as they evaluated 105 core needle biopsies in 2 phases: first unaided by AI, and then assisted by Paige AI after a washout period. They found that pathologists requested 40% fewer second reads when assisted by Paige Prostate AI, as well as 20% fewer IHC studies. The AI also allowed pathologists to decrease the time required for reading and reporting each slide by 21.94%. The study concluded that “the synergic usage of Paige Prostate contributes to a significant decrease in IHC studies, second opinion requests, and time for reporting while maintaining highly accurate diagnostic standards.”1
Ultimately, combining the strengths of both pathologists and AI creates the most robust possible diagnostic test.. In keeping patients at the forefront of clinical decision-making, labs should adopt AI in a way that best supports their individual pathologists and sets them up for success. Paige is the perfect partner to support labs in this mission, and our purpose-driven, clinical-grade AI can complement your pathologists as they continue the tireless fight against cancer, and create a better world for patients everywhere.
To learn more about how Paige’s AI can support your team, schedule a demo.
—
References
1Eloy C, Marques A, Pinto J, et al. Artificial intelligence–assisted cancer diagnosis improves the efficiency of pathologists in prostatic biopsies. Virchows Arch. 2023. doi: 10.1007/s00428-023-03518-5
In the United States, Paige Prostate Detect (DEN200080) is approved for clinical use with Philips Ultrafast Scanner. Otherwise, all Paige Prostate Suite applications use are limited to Research Use Only and not for use in diagnostic procedures.
*Paige Prostate has a sensitivity of 99.7% and a specificity of 99.3%, according to Perincheri S, Levi AW, Celli R, et al. An independent assessment of an artificial intelligence system for prostate cancer detection shows strong diagnostic accuracy. Mod Pathol. 2021;34(8):1588-1595. doi: 10.1038/s41379-021-00794-x. As of March 1 2022, Paige Prostate was updated to Paige Prostate Suite, under which Paige Prostate Detect and Paige Prostate Grade and Quantify became two different products. Both Paige Prostate Detect and Paige Prostate Grade and Quantify were used in this study.
